2 * Copyright 2015 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
25 #include <drm/drm_auth.h>
27 #include "amdgpu_sched.h"
28 #include "amdgpu_ras.h"
30 #define to_amdgpu_ctx_entity(e) \
31 container_of((e), struct amdgpu_ctx_entity, entity)
33 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
34 [AMDGPU_HW_IP_GFX] = 1,
35 [AMDGPU_HW_IP_COMPUTE] = 4,
36 [AMDGPU_HW_IP_DMA] = 2,
37 [AMDGPU_HW_IP_UVD] = 1,
38 [AMDGPU_HW_IP_VCE] = 1,
39 [AMDGPU_HW_IP_UVD_ENC] = 1,
40 [AMDGPU_HW_IP_VCN_DEC] = 1,
41 [AMDGPU_HW_IP_VCN_ENC] = 1,
42 [AMDGPU_HW_IP_VCN_JPEG] = 1,
45 static int amdgput_ctx_total_num_entities(void)
47 unsigned i, num_entities = 0;
49 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
50 num_entities += amdgpu_ctx_num_entities[i];
55 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
56 enum drm_sched_priority priority)
58 /* NORMAL and below are accessible by everyone */
59 if (priority <= DRM_SCHED_PRIORITY_NORMAL)
62 if (capable(CAP_SYS_NICE))
65 if (drm_is_current_master(filp))
71 static int amdgpu_ctx_init(struct amdgpu_device *adev,
72 enum drm_sched_priority priority,
73 struct drm_file *filp,
74 struct amdgpu_ctx *ctx)
76 unsigned num_entities = amdgput_ctx_total_num_entities();
80 if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
83 r = amdgpu_ctx_priority_permit(filp, priority);
87 memset(ctx, 0, sizeof(*ctx));
90 ctx->fences = kcalloc(amdgpu_sched_jobs * num_entities,
91 sizeof(struct dma_fence*), GFP_KERNEL);
95 ctx->entities[0] = kcalloc(num_entities,
96 sizeof(struct amdgpu_ctx_entity),
98 if (!ctx->entities[0]) {
100 goto error_free_fences;
103 for (i = 0; i < num_entities; ++i) {
104 struct amdgpu_ctx_entity *entity = &ctx->entities[0][i];
106 entity->sequence = 1;
107 entity->fences = &ctx->fences[amdgpu_sched_jobs * i];
109 for (i = 1; i < AMDGPU_HW_IP_NUM; ++i)
110 ctx->entities[i] = ctx->entities[i - 1] +
111 amdgpu_ctx_num_entities[i - 1];
113 kref_init(&ctx->refcount);
114 spin_lock_init(&ctx->ring_lock);
115 mutex_init(&ctx->lock);
117 ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
118 ctx->reset_counter_query = ctx->reset_counter;
119 ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
120 ctx->init_priority = priority;
121 ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
123 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
124 struct amdgpu_ring *rings[AMDGPU_MAX_RINGS];
125 struct drm_sched_rq *rqs[AMDGPU_MAX_RINGS];
127 unsigned num_rqs = 0;
130 case AMDGPU_HW_IP_GFX:
131 rings[0] = &adev->gfx.gfx_ring[0];
134 case AMDGPU_HW_IP_COMPUTE:
135 for (j = 0; j < adev->gfx.num_compute_rings; ++j)
136 rings[j] = &adev->gfx.compute_ring[j];
137 num_rings = adev->gfx.num_compute_rings;
139 case AMDGPU_HW_IP_DMA:
140 for (j = 0; j < adev->sdma.num_instances; ++j)
141 rings[j] = &adev->sdma.instance[j].ring;
142 num_rings = adev->sdma.num_instances;
144 case AMDGPU_HW_IP_UVD:
145 rings[0] = &adev->uvd.inst[0].ring;
148 case AMDGPU_HW_IP_VCE:
149 rings[0] = &adev->vce.ring[0];
152 case AMDGPU_HW_IP_UVD_ENC:
153 rings[0] = &adev->uvd.inst[0].ring_enc[0];
156 case AMDGPU_HW_IP_VCN_DEC:
157 rings[0] = &adev->vcn.ring_dec;
160 case AMDGPU_HW_IP_VCN_ENC:
161 rings[0] = &adev->vcn.ring_enc[0];
164 case AMDGPU_HW_IP_VCN_JPEG:
165 rings[0] = &adev->vcn.ring_jpeg;
170 for (j = 0; j < num_rings; ++j) {
174 rqs[num_rqs++] = &rings[j]->sched.sched_rq[priority];
177 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j)
178 r = drm_sched_entity_init(&ctx->entities[i][j].entity,
179 rqs, num_rqs, &ctx->guilty);
181 goto error_cleanup_entities;
186 error_cleanup_entities:
187 for (i = 0; i < num_entities; ++i)
188 drm_sched_entity_destroy(&ctx->entities[0][i].entity);
189 kfree(ctx->entities[0]);
197 static void amdgpu_ctx_fini(struct kref *ref)
199 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
200 unsigned num_entities = amdgput_ctx_total_num_entities();
201 struct amdgpu_device *adev = ctx->adev;
207 for (i = 0; i < num_entities; ++i)
208 for (j = 0; j < amdgpu_sched_jobs; ++j)
209 dma_fence_put(ctx->entities[0][i].fences[j]);
211 kfree(ctx->entities[0]);
213 mutex_destroy(&ctx->lock);
218 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
219 u32 ring, struct drm_sched_entity **entity)
221 if (hw_ip >= AMDGPU_HW_IP_NUM) {
222 DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
226 /* Right now all IPs have only one instance - multiple rings. */
228 DRM_DEBUG("invalid ip instance: %d\n", instance);
232 if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
233 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
237 *entity = &ctx->entities[hw_ip][ring].entity;
241 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
242 struct amdgpu_fpriv *fpriv,
243 struct drm_file *filp,
244 enum drm_sched_priority priority,
247 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
248 struct amdgpu_ctx *ctx;
251 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
255 mutex_lock(&mgr->lock);
256 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
258 mutex_unlock(&mgr->lock);
264 r = amdgpu_ctx_init(adev, priority, filp, ctx);
266 idr_remove(&mgr->ctx_handles, *id);
270 mutex_unlock(&mgr->lock);
274 static void amdgpu_ctx_do_release(struct kref *ref)
276 struct amdgpu_ctx *ctx;
277 unsigned num_entities;
280 ctx = container_of(ref, struct amdgpu_ctx, refcount);
283 for (i = 0; i < AMDGPU_HW_IP_NUM; i++)
284 num_entities += amdgpu_ctx_num_entities[i];
286 for (i = 0; i < num_entities; i++)
287 drm_sched_entity_destroy(&ctx->entities[0][i].entity);
289 amdgpu_ctx_fini(ref);
292 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
294 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
295 struct amdgpu_ctx *ctx;
297 mutex_lock(&mgr->lock);
298 ctx = idr_remove(&mgr->ctx_handles, id);
300 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
301 mutex_unlock(&mgr->lock);
302 return ctx ? 0 : -EINVAL;
305 static int amdgpu_ctx_query(struct amdgpu_device *adev,
306 struct amdgpu_fpriv *fpriv, uint32_t id,
307 union drm_amdgpu_ctx_out *out)
309 struct amdgpu_ctx *ctx;
310 struct amdgpu_ctx_mgr *mgr;
311 unsigned reset_counter;
316 mgr = &fpriv->ctx_mgr;
317 mutex_lock(&mgr->lock);
318 ctx = idr_find(&mgr->ctx_handles, id);
320 mutex_unlock(&mgr->lock);
324 /* TODO: these two are always zero */
325 out->state.flags = 0x0;
326 out->state.hangs = 0x0;
328 /* determine if a GPU reset has occured since the last call */
329 reset_counter = atomic_read(&adev->gpu_reset_counter);
330 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
331 if (ctx->reset_counter_query == reset_counter)
332 out->state.reset_status = AMDGPU_CTX_NO_RESET;
334 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
335 ctx->reset_counter_query = reset_counter;
337 mutex_unlock(&mgr->lock);
341 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
342 struct amdgpu_fpriv *fpriv, uint32_t id,
343 union drm_amdgpu_ctx_out *out)
345 struct amdgpu_ctx *ctx;
346 struct amdgpu_ctx_mgr *mgr;
347 uint32_t ras_counter;
352 mgr = &fpriv->ctx_mgr;
353 mutex_lock(&mgr->lock);
354 ctx = idr_find(&mgr->ctx_handles, id);
356 mutex_unlock(&mgr->lock);
360 out->state.flags = 0x0;
361 out->state.hangs = 0x0;
363 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
364 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
366 if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
367 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
369 if (atomic_read(&ctx->guilty))
370 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
373 ras_counter = amdgpu_ras_query_error_count(adev, false);
374 /*ras counter is monotonic increasing*/
375 if (ras_counter != ctx->ras_counter_ue) {
376 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
377 ctx->ras_counter_ue = ras_counter;
381 ras_counter = amdgpu_ras_query_error_count(adev, true);
382 if (ras_counter != ctx->ras_counter_ce) {
383 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
384 ctx->ras_counter_ce = ras_counter;
387 mutex_unlock(&mgr->lock);
391 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
392 struct drm_file *filp)
396 enum drm_sched_priority priority;
398 union drm_amdgpu_ctx *args = data;
399 struct amdgpu_device *adev = dev->dev_private;
400 struct amdgpu_fpriv *fpriv = filp->driver_priv;
403 id = args->in.ctx_id;
404 priority = amdgpu_to_sched_priority(args->in.priority);
406 /* For backwards compatibility reasons, we need to accept
407 * ioctls with garbage in the priority field */
408 if (priority == DRM_SCHED_PRIORITY_INVALID)
409 priority = DRM_SCHED_PRIORITY_NORMAL;
411 switch (args->in.op) {
412 case AMDGPU_CTX_OP_ALLOC_CTX:
413 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
414 args->out.alloc.ctx_id = id;
416 case AMDGPU_CTX_OP_FREE_CTX:
417 r = amdgpu_ctx_free(fpriv, id);
419 case AMDGPU_CTX_OP_QUERY_STATE:
420 r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
422 case AMDGPU_CTX_OP_QUERY_STATE2:
423 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
432 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
434 struct amdgpu_ctx *ctx;
435 struct amdgpu_ctx_mgr *mgr;
440 mgr = &fpriv->ctx_mgr;
442 mutex_lock(&mgr->lock);
443 ctx = idr_find(&mgr->ctx_handles, id);
445 kref_get(&ctx->refcount);
446 mutex_unlock(&mgr->lock);
450 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
455 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
459 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
460 struct drm_sched_entity *entity,
461 struct dma_fence *fence, uint64_t* handle)
463 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
464 uint64_t seq = centity->sequence;
465 struct dma_fence *other = NULL;
468 idx = seq & (amdgpu_sched_jobs - 1);
469 other = centity->fences[idx];
471 BUG_ON(!dma_fence_is_signaled(other));
473 dma_fence_get(fence);
475 spin_lock(&ctx->ring_lock);
476 centity->fences[idx] = fence;
478 spin_unlock(&ctx->ring_lock);
480 dma_fence_put(other);
485 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
486 struct drm_sched_entity *entity,
489 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
490 struct dma_fence *fence;
492 spin_lock(&ctx->ring_lock);
495 seq = centity->sequence - 1;
497 if (seq >= centity->sequence) {
498 spin_unlock(&ctx->ring_lock);
499 return ERR_PTR(-EINVAL);
503 if (seq + amdgpu_sched_jobs < centity->sequence) {
504 spin_unlock(&ctx->ring_lock);
508 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
509 spin_unlock(&ctx->ring_lock);
514 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
515 enum drm_sched_priority priority)
517 unsigned num_entities = amdgput_ctx_total_num_entities();
518 enum drm_sched_priority ctx_prio;
521 ctx->override_priority = priority;
523 ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
524 ctx->init_priority : ctx->override_priority;
526 for (i = 0; i < num_entities; i++) {
527 struct drm_sched_entity *entity = &ctx->entities[0][i].entity;
529 drm_sched_entity_set_priority(entity, ctx_prio);
533 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
534 struct drm_sched_entity *entity)
536 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
537 unsigned idx = centity->sequence & (amdgpu_sched_jobs - 1);
538 struct dma_fence *other = centity->fences[idx];
542 r = dma_fence_wait(other, true);
544 if (r != -ERESTARTSYS)
545 DRM_ERROR("Error (%ld) waiting for fence!\n", r);
554 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
556 mutex_init(&mgr->lock);
557 idr_init(&mgr->ctx_handles);
560 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
562 unsigned num_entities = amdgput_ctx_total_num_entities();
563 struct amdgpu_ctx *ctx;
567 idp = &mgr->ctx_handles;
569 mutex_lock(&mgr->lock);
570 idr_for_each_entry(idp, ctx, id) {
571 for (i = 0; i < num_entities; i++) {
572 struct drm_sched_entity *entity;
574 entity = &ctx->entities[0][i].entity;
575 timeout = drm_sched_entity_flush(entity, timeout);
578 mutex_unlock(&mgr->lock);
582 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
584 unsigned num_entities = amdgput_ctx_total_num_entities();
585 struct amdgpu_ctx *ctx;
589 idp = &mgr->ctx_handles;
591 idr_for_each_entry(idp, ctx, id) {
592 if (kref_read(&ctx->refcount) != 1) {
593 DRM_ERROR("ctx %p is still alive\n", ctx);
597 for (i = 0; i < num_entities; i++)
598 drm_sched_entity_fini(&ctx->entities[0][i].entity);
602 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
604 struct amdgpu_ctx *ctx;
608 amdgpu_ctx_mgr_entity_fini(mgr);
610 idp = &mgr->ctx_handles;
612 idr_for_each_entry(idp, ctx, id) {
613 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
614 DRM_ERROR("ctx %p is still alive\n", ctx);
617 idr_destroy(&mgr->ctx_handles);
618 mutex_destroy(&mgr->lock);